Integrated Approach for the Detection of Learning Styles & Affective States
Farman Ali Khan1, Sabine Graf2, Edgar R. Weippl1, A Min Tjoa1
1
Institute of Software Technology and Interactive Systems
Vienna University of Technology
2
Graduate Institute of Learning and Instruction, National Central University, Taiwan
farman@ifs.tuwien.ac.at, sabine.graf@ieee.org,weippl@securityresearch.at, tjoa@ifs.tuwien.ac.at
Abstract: Detecting the needs of the learner is a challenging task for systems addressing
issues related to either adaptivity or personalization. In this paper we present a student-driven
learning and assessment tool along with a tool for detecting/calculating learning styles and
affective states. The learning and assessment tool enables learning, while the
detection/calculation tool detects learning styles and affective states from the behavior of
learners. The resulting behavior is then reflected in the database. This method does not
require learners to fill out a lengthy questionnaire. Also, there is no need to monitor
observational cues, such as gesture, posture, conversation etc. of the learner; the
detection/calculation tool automatically identifies learning styles and affective states.
Introduction
Modeling the features of the learner is pertinent for systems providing methods to deal with either adaptivity or
personalization issues (Graf et al. 2006). For that reason, the user model is a main component of any adaptive
system. The user model presents information about an individual user, allowing adaptive systems to use this
information to provide adaptation effects (Brusilovsky 2007). When building an adaptive system, adaptation is
usually based on a user’s knowledge, goals, and preferences. Adapting to a user’s personality traits or his/her
affective states, such as motivation and emotion, are rarely considered (Dautenhahn 2002).
(Graf and Kinshuk 2006) emphasized the fact that the needs of the learners must first be known before
adaptivity can be provided. (Bruisilovsky 1996) has mentioned two different approaches for obtaining
information about a learner’s needs: the collaborative and the automatic student modeling approach. In the
collaborative modeling approach, learners provide explicit information about themselves by filling out a
questionnaire, whereas in the automatic modeling approach, the system monitors the actions and behavior of
learners and infer their needs automatically while they are learning/working within the system.
In systems that provide adaptation based on the affective state of the learner, observational cues such as gesture,
posture, conversation, etc. are used. According to (Cocea 2006), it is very difficult for adaptive systems to
process these kinds of observational cues. That is why most research attention is shifting towards systems that
automatically process motivation/emotion cues as a means to assess affective states.
With traditional teaching methods and high student/teacher ratios, a teacher faces great obstacles in the
classroom. Traditionally, teachers deliver the content and students learn it. With this methodology, teachers are
usually not able to respond to the individual needs of students. Furthermore, due to the large numbers of
students in a classroom, teachers are not able to focus on each individual student. For grading purposes they
normally conduct an exam at the end of the course. However, despite the high number of students in a
classroom, experienced teachers usually observe, recognize and address the learning style and affective state of
the learners. The skilled teacher takes suitable action to positively impact learning. But the question is what
these experienced teachers “see” and how they arrive at a course of action and whether this action leads the
learner to a productive path.
We believe that accurately identifying learning styles and affective states is a critical step in understanding how
to improve the learning process. We also believe that computers will soon be capable of recognizing human
behavior and producing accurate assessments regarding learning styles and affective states.
3
Previous Research and Our Perspective
Previous Research
Related work deals with the identification of learning styles in adaptive systems, such as ILASH (Peña 2002),
MASPLANG (Marzo et al 2003), and CS383(Carver 1999). These systems use collaborative student modeling
or a self reported informational approach for detection of learning styles. This approach has at least two
potential downfalls (Shute 2007).
1.
2.
First, learners can provide inaccurate data due to a lack of knowledge about their own characteristics either
accidentally or purposely, due to privacy concerns.
Second, during the online learning process, completing the questionnaire can be time consuming, which
might frustrate learners and lead them to provide invalid data in order to arrive at the content more quickly.
Only a few other systems adopted automatic modeling approaches such as Arthur (Gilbert 1999), which is a
Web-based instruction system. In an automatic student modeling approach there is no additional effort necessary
on the part of learners in order to obtain information about their learning styles. Learners simply interact with
the system while learning and the system infers their learning styles and accommodates their needs. The
information obtained in this way is free from uncertainty. DELES (Graf 2007) detects learning styles from the
learners’ different patterns of behavior, such as the number of visits in a forum, the number of times they
participate in a chat, the number of postings in a forum and the number of visits, and the time a learner takes to
deal with exercises.
Related work regarding identification of affective states comes either from observation cues, such as
Itspoke(D’Mello 2006) or Mota(Kapoor et al 2004) or through interaction with the system, such as Vfts(Roll
2005).The affective state detection through observational cues has one drawback, that is, once learners become
aware that their learning activities are being monitored, there is a chance that they will become nervous or
anxious or even depressed before beginning their regular learning activity and will perhaps continue to maintain
this emotional state throughout the learning experience.
Our Perspective
We propose a simple architecture that is easy for learners to use, for detecting learning styles and affective
states. Our approach integrates information about learning styles and affective states, such as
motivation/emotions, in order to enable educational systems to provide personalized and more efficient
adaptation based on the features shown in Figure 1.
Our approach aims at inferring the learning styles and affective states (i.e., motivational/emotional states) from
the behavior of the learners instead of using a questionnaire to ask the learners about their preferences for
learning styles and observational cues for affective states. The proposed learning and assessment
tool/framework tracks the learners’ knowledge progress, provides an assessment, and supports the learner by
offering hints, help, and feedback. The proposed learning styles and affective states tool not only calculates the
learning style from patterns of behavior but it also predicts certain components of the affective state i.e.
motivational/emotional state of each individual user while interacting with the system.
Moreover, we propose an indirect approach to motivate/ encourage students to use the student driven learning
and assessment tool by introducing classroom quizzes, as classroom quizzes have a great role in motivating
students to learn (Brusilovsky 2005). This can be accomplished by introducing 8 to 10 classroom quizzes per
semester/course. During a quiz the student will be asked to answer 10 questions within 10 minutes related to
topics discussed during the most recent last lecture.
This approach not only motivates, but engages the student to use student driven learning and assessment tools
for learning. Thus the students who worked with the learning and assessment tools after each lecture will be able
to answer most of the questions in the classroom. These approaches enable the categorization into a database of
behavioral patterns, calculation of learning styles, and prediction of affective states for a majority of students
Figure 1: User Model/Profile
Learning and Assessment Tools Architecture
In this section we present the student-driven learning and assessment tool, in which an initial learning activity
takes place followed by an assessment. The architecture is shown in Figure 2
Figure 2: Learning and Assessment
The learning and assessment tool first offers the learner a “learning” or lesson for learning. The learning format
is text with graphics/diagrams, as applicable. In the learning activity/process chapters are presented. Learners
will have to select one of the chapters presented. Each chapter consists of a number of topics. Upon
selecting/clicking a topic, its contents are presented in a page/slide format. This action of selection/clicking is
recorded by the system i.e. it means that the learner has visited that page/slide. The learner has the option to
select the next topic randomly instead of following the sequential order. When the learner finishes reading the
chapter, the system will check whether the learner has visited 50 percent of the topics of the chapter/module. If
he/she has done so, the system will ask the learner if he/she wishes to move ahead to the next exercise or
complete the reading of the existing contents. If the learner selects the option to go ahead, an exercise is offered.
Otherwise, the system switches back to reading the content section again. If the learner has not read 50 percent
of the contents of the chapter/module, the system will recommend the learner, to continue reading the remaining
content. The proposed tool/framework supports and motivates the learners’ during the assessment process in the
form of offering hints and help. Each assessment exercise consists of two parts. The first part comprises multiple
choice questions. At this point, a help window is available. The learner can use the help window to reread the
applicable topic. Here the learner will be able to avail the help facility a fixed number of times i.e. 3 times while
attempting the first part of exercise. After finishing the first part, the second part of the exercise requires filling
in the blanks. In this second part, a relevant hint is available. If learners do not understand the particular
question, they can look for a hint; often a more precise and simple explanation. Here again the learner will be
able to avail the hint facility a fixed number of times. At the end of each exercise, results are validated. If
incorrect, the correct solutions are described. The system continues in this manner for each chapter. The help
screen and the hints are provided in order to motivate the learner and prevent giving up or dropping out. The
restriction of offering the help and hint window/facility only a fixed number of times has been introduced in
order to avoid “help and hint abuse” i.e. those learners who quickly and repeatedly ask for help and hint instead
of attempting the exercise on his or her own. When the learner has finished all the learning and assessment
modules/lessons, a final miscellaneous exercise is offered. These final questions are based on all of the prior
learning modules/lessons. No help or hint is offered, in order to rate improvement.
Learning Style Models
There are several theories pertaining to learning styles, such as Felder-Silverman learning style model (FSLSM)
(Felder 1988), Dunn & Dunn Learning style model (Dunn 2001), and Kolb’s Learning style model(Kolb 1984),
etc. While other learning style models classify learners in few groups, the Felder-Silverman learning style model
categorizes the learning styles of learners based on four dimensions (active/reflective, sensing/intuitive,
visual/verbal and sequential/global), so that each learner has a preference for each of these four dimensions.
That is why FSLSM, shown in Fig. 3, is most appropriate for educational systems; and our work is based on it.
S.No Dimension
Definition
1
Verbal
Require written or spoken
Visual
Remember what they have seen
Sequential
Learn in linear steps
Global
Holistic or learn in large leaps
Active
Learn by trying things
Reflective
Learn by thinking things out
Sensing
Learn concrete material and tend
to be practical
Intuitive
learners
Learn concepts
2
3
4
Figure. 3. Felder-Silverman learning style model
Affective States in E-learning
In most of the e-Learning systems, including adaptive systems, attention is either towards knowledge acquisition
or cognitive processing. When building a system, affective states such as motivation and emotion, are
considered only in terms of how the content is structured and presented. To make learning efficient and to
deliver personalized content, adaptive systems are based on models of a user’s goals, knowledge, and
preferences. Thus, a user model that integrates the cognitive processes and motivational states would lead to
more efficient and personalized adaptation (Cocea 2007).Transforming a non-affect sensitive system into a
system that is responsive to a user’s affective states requires the modeling of a cycle known as the affective
loop. The affective loop encompasses detection of a user’s affective states, appropriate actions selection for
decision making, and the synthesis of appropriate affective state by the system (D'Mello 2008).
According to Weimin(2007), affection influences the learning performance and decision making. This means
that students who become caught in affective states such as anger or depression do not absorb information
efficiently. From this, it can be inferred that a user’s affective state has a major role in improving the
effectiveness of e-learning.
Patterns of Behavior
Detecting learning styles and affective states occurs by detecting patterns that indicate a preference for a specific
dimension and also for a specific affective state. We focused on commonly used features, such as content
objects and exercises. The patterns of behavior considered for detecting learning styles and affective states are:
1.
2.
3.
Time spent learning/reading each
activity i.e. contents.
Time spent learning/reading contents
containing graphics/diagram.
Time spent solving exercises.
4.
5.
Number of incorrect answers in each
exercise related to graphics/diagrams.
Number of incorrect answers in each
exercise.
6.
Number of incorrect answers in final
12. Attempted order to answer questions
miscellaneous exercise.
in exercise (sequential/ random).
7. Number of changes to the exercise
13. Skipping pages/slide content sections.
answer.
14. Mouse clicking through the interface.
8. Number of questions left unattended
15. How often and how insistent the
in each exercise per session.
learner looks for help.
9. Number of independent sessions of
16. How often and how insistent the
work with the system.
learner looks for hints.
10. Number of solved exercises.
11. How many times the student has
reviewed the topic/contents.
These patterns of behavior are also called observable variables. Each of these patterns gives an indication related
to a learning style dimension and an affective state.
Learning Styles from Patterns of Behavior
1. Active/Reflective Learners
According to FSLSM, active learners prefer to process information actively by doing something with the
learned material, such as discussing it, explaining it, or testing it. We can therefore assume that the following
behavior provides us with hints about the learner’s preference for an active learning style: Low number of
sessions to complete the learning process, less time taken to complete the reading/learning content as compared
with the time taken to solve the exercises, low number of times the learner looks for help, low number of times
the learner looks for hints and does so less insistently, high number of performed exercises.
According to FSLSM, reflective learners prefer to think about and reflect on the learning material. Moreover
they prefer to work alone. We can therefore assume that a situation that provides little or no opportunity to think
about the information presented, reflective learners do not learn properly. Furthermore we can also assume that
spending more time on learning/reading contents provides us with hints about the learner’s preference for a
reflective learning style.
2. Sensing/Intuitive Learners
According to FSLSM, sensing learners like to learn facts and concrete learning material. Sensing learners also
dislike challenges/complications and like to solve problems through well-established methods. Sensing learners
usually work carefully and slowly. We can therefore assume that sensing behavior of learners can be predicted
through: Number of times the student has reviewed the topic/contents, Number of times he or she changes the
exercise answers, Time spent on solving the exercise.
According to FSLSM, intuitive learners welcome challenges and are bored by details. Intuitive learners tend to
work faster. Furthermore, they like innovation and dislike repetition. We can therefore assume that a learner
who does not revise his exam /assessment test tends to have an intuitive learning style. Moreover, high the
number of solved exercises and low the time spent on solving each exercise also reveals the learner’s aptitude
for intuitive behavior.
3. Visual/Verbal Learners
According to FSLSM, visual learners remember best what they see: demonstrations, films, diagram, and
pictures; and create appropriate mental images of it. We can therefore assume that spending less time on
learning/reading contents that contains graphics/diagrams and then performing well in exercises on questions
related to such contents gives an indication of a learner’s visual aptitude.
According to FSLSM, verbal learners prefer words either oral or written, as their method of learning. We can
therefore assume that a great deal of time on learning/reading contents that contain graphics/diagrams and
showing poor performance in exercise on questions related to those contents gives us an indication of his/her
verbal learning aptitude. Conversely, performing well in exercises on questions related to the textual content
further strengthens the indication of his/her verbal learning style.
4. Sequential/Global Learners
According to FSLSM, sequential learners not only explore the course material sequentially, in detail but also
follow linear reasoning processes when solving problems. We can therefore assume that not skipping
pages/slide content sections can give a hint for a sequential learning style, indicating sequential behavior in
exploring the course material. Moreover, the sequential selection/clicking order of questions in exercises also
hint at a learner’s aptitude for sequential behavior.
According to FSLSM, global learners are not interested in obtaining details of the contents being presented but
instead, like to get an overview of the contents. Using this way of learning they get the big picture and build
their own cognitive map of the contents. We can therefore assume that skipping pages/slide content sections can
give a hint for a global learning style, indicating global behavior in exploring the course material. Moreover, the
random selection/clicking order of questions in exercises also hint at a learner’s aptitude for global behavior.
Affective States from Patterns of Behavior
(Qu 2005) described several factors including confidence, confusion, and effort that influence learners’ affective
states, e.g., motivation. These factors have been said to be important following the background tutor studies.
(Vicente 2002) described in their motivational model an independence that is relevant to the challenge and also
relates to interpersonal motivational factors, such as cooperation, competition, and recognition.
The variables in our approach are: confidence, confusion, effort, and independence. These variables represent
characteristics of the students that relate to the contents and exercises. Following the Qu model and Vicente
motivational model, affective states predicted from the patterns of behavior are explained below.
1. Confidence
According to Qu’s model for pedagogical agents, the confidence of a learner in a learning environment is
defined by the confidence of the learner in solving problems. We can therefore assume that the following
behavior provides us with hints about the student’s confidence distinguishing between high, normal, and low
confidence i.e. High number of correct answers in each exercise, High number of solved exercises. A learner’s
confidence can also be inferred from mouse movement/clicking throughout the interface. This movement /
clicking of the learner also confirm that the learner was paying attention to the task / exercise answers. If the
task is performed quickly and is also performed well, it can be inferred that the learner was confident. If the
learner performed a task quickly, it can also mean a lack of interest. But the combination of other pieces of
evidence, such as that the learner was interested in performing the task and also he /she performed well, leads to
the assumption that in such cases, quick performance is due to high confidence.
2. Confusion
According to Qu’s model for pedagogical agents a learner in highly confused state is most likely to be stuck or
frustrated. We can therefore assume that the following behavior provides us with hints about the student’s
confusion i.e. High number of incorrect answers in each exercise, Low number of solved exercises. Also the
state of confusion can be inferred from how many times the learner has asked for help or hint in solving the
exercise. Moreover the number of questions left unattended in each exercise per session i.e. not completing the
exercise and leaving the session abruptly, gives an indication of the learner’s state of confusion.
3. Effort
According to Qu’s model for pedagogical agents, a learner’s effort can be seen in the amount of time the learner
spends on performing tasks. We can therefore assume that the following behavior provides us with information
related to the student’s effort: The amount of time the learner spends on performing tasks, such as the total time
taken reading the contents and solving exercises, predicts a learner’s effort. Effort is a fundamental indicator of
a learner’s motivation. The exertion of effort is considered by human tutors as a positive parameter and they
praise learners for extending effort even when they are not successful.
4. Independence
According to Vicente’s motivational model, a learner’s independence is judged from the efforts he or she makes
to solve the problems on his or her own without asking for help from others even if the student encounters some
difficulty. The greater the numbers of sessions of work with the system and the greater the time spent on solving
exercises gives an indication that learner was facing difficulty. Nonetheless, performing well on exercises
without getting hints or help from the system at any stage offers an indication of the learner’s independence.
Learning Styles and Affective States Tools Architecture
In this section we present a tool/framework for detecting learning styles and affective states, such as motivation
and emotion based on a learner’s behavior. The tools architecture is divided into two components as shown in
Figure 4. The data in the database is stored in an event-based way, which simplifies the data extraction process
from the database. The data extraction component is shown in figure 4.
Figure 4. Detecting learning styles and affective states [Figure taken from Graf(2007)].
Calculation of Learning Styles and Affective States
Learning styles and affective states are calculated from the raw data extracted by the data extraction component.
Learning styles are calculated in this way for each dimension based on the ordered data. Ordered data for each
pattern can take the values +1, 0,-1, indicating, e.g. a balanced, moderate, or strong state. Affective states are
also calculated for each aspect based on the ordered data. Ordered data for each pattern can also take the values
+1, 0,-1 indicating a low, normal, or high state. The proposed tool recommends thresholds, which can be
changed if necessary. The idea of determining thresholds for learning styles and affective states derived from
(García et al. 2007) dealing with the usage of respective features.
Calculation of learning styles is based on the approach adopted in the ILS(Index of Learning Styles), which is a
questionnaire for identifying learning styles according to FSLSM (Felder 2005). According to this approach, the
ordered data relevant to each dimension are summarized. Each dimension result is converted to a 3-item scale
indicating, for example a balanced, moderate, or strong preference. Similarly, in the calculation of affective
states, ordered data relevant to each aspect of the affective state are summarized. Each aspect of the affective
state result is converted to a 3-item scale indicating low, normal, and high states.
Conclusion and Future Work
In this paper, we presented a student-driven learning and assessment tool along with learning styles and
affective states detection/calculation tool. The learning and assessment tool enables both learning and
assessment. The learning styles and affective states detection/calculation tool detects learning styles and
affective states from the behavior of learners reflected in the database by the learning and assessment tool.
Applying the proposed approach enables us to detect learning styles and affective states from the behavior of
learners. Therefore, to obtain information about the learners’ learning styles, learners’ do not need to fill out a
questionnaire. Also, it is not necessary to monitor observational cues, such as gesture, posture, conversation
patterns, etc., of the learner to obtain information about a learner’s affective states. The detection/calculation can
be done automatically by the learning styles and affective states detection/calculation tool.
Future work deals with investigating the suitability of the selected patterns for the detection of learning styles
and affective states, i.e., correlation between patterns of behavior and learning styles, affective states.
Furthermore, we also plan to develop the proposed tool for the detection of learning styles and affective states.
References
Brusilovsky, P.(1996). Methods and techniques of adaptive hypermedia, Journal of User Modelling and User
Adaptation Interaction, 6( 2-3), 87-129.
Brusilovsky, P., Kobsa, A., and Nejdl, W. (Eds).(2007). The Adaptive Web: Methods and Strategies of Web Personalization.
LNCS Springer Berlin / Heidelberg.
Brusilovsky, P., Sonovsky, S. (2005).Engaging students to work with self-assessment questions: A study of two approaches.
Proceedings of 10th Annual Conference on Innovation and Technology in Computer Science Education (ITiCSE'2005),
ACM Press, 251-255 Monte de Caparica, Portugal.
Carver, C.A., Howard, R.A., and Lane, W.D.(1999). Addressing different learning styles through course hypermedia. IEEE
Transactions on Education, 42 (1), 33-38.
Cocea, M.(2006). Assessment of motivation in online learning environments. Adaptive Hypermedia and Adaptive Web-Based
Systems, lncs Springer-Verlag Berlin / Heidelberg .
Cocea, M., Weibelzahl, S. (2007). Eliciting motivation knowledge from log files towards motivation diagnosis for Adaptive
Systems. User Modeling 2007 LNCS Springer Berlin / Heidelberg.
Dautenhahn, K., Waern, A. (2002). Book Review Rosalind Picard: Affective Computing, Journal of User Modeling and
User-Adapted Interaction, 12(1), 85-89.
D'Mello, S. K., Craig, S.D., Witherspoon, A., McDaniel, B., and Graesser, A., (2008). Automatic Detection of Learner’s
Affect from Conversational Cues. Journal of User Modeling and User-Adapted Interaction, 18(1-2), 45-80.
D'Mello, S. K., Craig, S.D., Sullins, J, and Graesser, A.C.(2006). Predicting affective states from autotutor dialogue.
Available at http://www.autotutor.org/publications/newpapers/dmello-dialogue-ijaied-06.pdf.
Dunn, R., Thies, A. P., & Honigsfeld(2001). A. Synthesis of the Dunn and Dunn learning-style model research: Analysis
from a neuropsychological perspective. NY: St. John’s University’s Center for the Study of Learning and Teaching Styles.
Felder, R.M., Silverman, L.K.,(1988). Learning and Teaching Styles in Engineering Education, available at
http://www4.ncsu.edu/unity/lockers/users/f/felder/public/Papers/LS-1988.pdf,
Felder, R.M., Spurlin, J.(2005). Applications, Reliability, and Validity of the Index of Learning Styles Int. Journal. of Engg .
Education. 21(1), 103-112.
García, P., Amandi, A., Schiaffino, S., Campo, M.(2007). Evaluating Bayesian Networks' Precision for Detecting Students'
Learning Styles. Journal of Computers & Education, 49( 3), 794-808.
Gilbert, J.E., Han, C.Y.(1999). Adapting Instruction in Search of a Significant Difference, Journal of Network and Computer
Applications, 22(3), 149-160.
Graf, S.(2007). Adaptivity in Learning Management Systems Focusing on Learning Styles. Ph.D Thesis, Vienna University
of Technology .
Graf, S., Kinshuk.(2006). An Approach for Detecting Learning Styles in Learning Management Systems, Proceedings of the
International Conference on Advances Learning Technologies (ICALT 06), 161-163.
Graf, S., Lin, T., Jeffrey, L., and Kinshuk.(2006). An Exploratory Study of the Relationship between Learning Styles and
Cognitive Traits, Proceedings of the European Conference on Technology Enhanced Learning (EC-TEL 06), 470-475.
Kapoor, A., Picard, R.W., and Ivanov, Y.(2004). Probabilistic Combination of Multiple Modalities to Detect Interest.
Proceedings of the Pattern Recognition, 17th International Conference on (ICPR'04) Volume 3. 969 – 972.
Kolb, D.A.(1984). Experiential learning: experience as the source of learning and development. Englewood Cliffs, New
Jersey: Prentice Hall.
Marzo, J.L., Peña, C.I., Aguilar, M., Palencia, X., Alemany, A., Vallès, M., Johè, A.(2003). Adaptive multiagent system for a
web-based tutoring environment. Available at http://bcds.udg.es/papers/agentcities-final-report.pdf.
Peña, C.I., Marzo, J.L., and de la Rosa, J.L.(2002). Intelligent Agents in a Teaching and Learning Environment on the Web.
Proceedings of the International Conference on Advanced Learning Technologies.
Qu, L., Wang, N., Johnson, W.L.(2005).Using Learner Focus of Attention to Detect Learner Motivation Factors In:
Proceedings of 10th Annual Conference on User Modeling, 70-73.
Roll, I., Baker, R.S., Aleven, V., McLaren, B.M., and Koedinger, K.R. (2005).Modeling Students’ Metacognitive Errors in
Two Intelligent Tutoring Systems: UM 2005. LNCS , Springer Berlin / Heidelberg.
Shute,
V.J.,
Zapata-Rivera,
D.(2007)
Adaptive
http://www.etsliteracy.net/Media/Research/pdf/RR-07-05.pdf
Technologies
,Research
Report.
Available
at
Vicente, A.D., Pain, H.(2002) Informing the detection of the students' motivational state: an empirical study, Proceedings of
the Sixth International Conference on Intelligent Tutoring Systems, 933-943.
Weimin, X., Wenhong, X.(2007). E-Learning Assistant System Based on Virtual Human Interaction Technology, ICCS
2007, LNCS, Springer Berlin / Heidelberg.